it's moving fast. now i have started a process of using the known 'correct' translations as "feedback" for the stuff that was broken like the above, but still coherent enough you can see how it came to it.
this revealed oscillations over a small set of two texts line by line in each language, starting with bulgarian/english. the oscillations pointed to where the texts that it had in it were not enough to fixate the nodes in the lattice that marked each word and how it travels a short path from one language in a sentence, to another language, with each point forming a structure that gave its analogs. that basic principle was shown in the second generation.
there is probably another distinction i need to find, but i'm thinking that the different word orders may just resolve themselves with enough. with this new metric to determine how well the lattice has stabilised, probably there will be one more new feature of the patterns that emerges that closes the loop and it's a full text translation matrix.
it's tiny. it takes like 10 minutes to eat the bible and the tao te ching. it's feeding on subtitles now, currently focusing on prefecting this one translation pair. but once there is a foundation, english is in the center of all of them. after that, translating between the languages that are now arrayed in a star shape around english - the paths walk from one side of the star to the other side. all it needs to grow is enough line by line patterns that you feed sentence at a time one language, then the other. they seek their fit, new words land in new places, and the existing ones where a subtle distinction gave a different word... the nodes shift and change their distances so the short path to translate that sentence is now the correct one.
the most remarkable thing was figuring out that you can expand the number of nodes in the tree without rewriting anything, you just put a new number on a coordinate that effects a scaling operation that changes what was x, y, z to a new position that reflects a deeper mesh that more things can slot into. zero memory cost, traversal is converging rapidly towards optimization, and the data size is retardedly small. probably i will have 20 languages in under 100mb of lattice, a perfect translator.